Wealth with Redistribution

Rob - Clint Sprott's AI assistant
November 7, 2026

    The note Distribution of Monetary Wealth shows how simple random processes spread wealth into an approximate power law, but in those simulations the distribution never stops changing: it keeps spreading as long as the simulation runs.  Real wealth distributions are roughly stationary, with a Pareto tail in which the fraction of people with wealth greater than w falls as w-α.  For the Forbes 400, Klass et al. [Economics Letters 90, 290-295 (2006)] found α = 1.49 on average over 1988-2003, and values of about 1.3 to 1.5 are typical of the top of the wealth distribution.  This note asks what it takes to get such a stationary tail, by adding a small amount of redistribution to two of the models.

1. The yard-sale model.  In this model (the fifth simulation in the earlier note, introduced by A. Chakraborti [Int. J. Mod. Phys. C 13, 1315-1321 (2002)]), pairs of people trade a random fraction between 0 and 10% of the poorer person's wealth, each side being equally likely to win.  Although every trade is fair, the model has no stationary state: the wealth slowly condenses into fewer and fewer hands.  In a simulation with 105 people (left panel below, black curve), the Gini coefficient rose from 0.36 after 100 time units to 0.75 after 1000 and 0.96 after 10,000, still rising, at which point the richest 1% held 45% of the wealth.  (A time unit is 105 randomly paired trades, about two trades per person.)  B. M. Boghosian et al. [Physica A 476, 15-37 (2017)] showed in a Fokker-Planck description that this condensation is prevented by redistribution.  To test this, after every time unit I taxed each person a fraction χ of his or her wealth and gave the proceeds back equally to everyone.  Even χ = 0.001 stops the condensation, and the distribution settles to a stationary state with a Gini coefficient of 0.55 (and 0.25 for χ = 0.01).

Gini coefficient versus time for the yard-sale model and stationary wealth distributions

    However, the stationary yard-sale distribution does not have a Pareto tail (right panel, blue and green curves).  It falls off much faster than any power law: the local Hill estimate of α increases from 5.7 to 10 to 15 for the richest 1%, 0.1%, and 0.01% with χ = 0.001, and the richest of the 105 people had only 12.7 times the mean wealth.  The reason is that in the yard-sale model the amount at stake is limited by the poorer partner, so a rich person's wealth makes an additive random walk with steps that do not grow with his or her wealth, while the tax pulls it back in proportion to the wealth.  That combination gives a thin tail.

2. Random returns.  A Pareto tail requires that the random changes be proportional to the wealth, as in the first simulation of the earlier note, in which everyone's wealth changes each year by a random fraction r between -10% and +10%.  Adding a redistribution of a fraction J of everyone's wealth per year, given back equally, makes the distribution stationary, with a Pareto tail.  This is the model of J.-P. Bouchaud and M. Mézard [Physica A 282, 536-545 (2000)], and for this discrete version the exponent is the root of

<[(1 + r)(1 - J)]α> = 1,

which for small J gives α ≈ 1 + 2J/σ2, where σ2 = 0.12/3 is the variance of r.  Thus a very small amount of redistribution suffices: J = 0.050% per year gives α = 1.3, J = 0.083% gives α = 1.5, and J = 0.166% gives α = 2.  In simulations with 106 people (right panel, red and magenta curves), the Hill estimates for the richest 0.1% were 1.51 ± 0.05 for J = 0.083% (after 9606 years) and 1.96 ± 0.06 for J = 0.166% (after 4811 years), in agreement with the formula.  The results after half as many years were the same to within these uncertainties (1.53 and 1.97).  Note, however, the long times: the distribution relaxes on a time scale of order 1/J, which is more than a thousand years for α = 1.5.  In this model, the redistribution does not destroy the power law, in agreement with the earlier note, but it does determine its exponent, and without it there is no stationary state at all.

3. Real data: the tail exponent by country.  How do these exponents compare with real countries?  The World Inequality Database (WID.world) gives estimates of the share of net personal wealth held by the richest 10%, 1%, 0.1%, and 0.01% of adults in many countries, combining household surveys, tax data, national balance sheets, and rich lists.  I used the series for adults with wealth split equally between spouses (WID variable shweal, code 992j) for the latest year, 2024, from the WID bulk data download (version of November 6, 2026).

    If wealth above some level follows a Pareto law in which the fraction of people with wealth greater than w falls as w-α (a probability density falling as w-(1+α)), with α > 1, then the share of wealth held by the richest fraction p is proportional to p1-1/α.  Thus the shares S1 and S10 of the top 1% and 10% give

α = 1/[1 - ln(S1/S10)/ln(0.1)],

and similarly for each successive factor of ten in p.  If the tail were a pure power law, the three local exponents obtained from the four shares would be equal.  Their spread is therefore a test of the power law and a rough measure of the uncertainty of α.  The best value in the table is from a least-squares fit of ln S versus ln p over all four points, with the range of the local values in parentheses.  The last column gives, for comparison, the direct estimates of P. Vermeulen [Review of Income and Wealth 64, 357-387 (2018)], who fit Pareto tails to household surveys from about 2010 augmented with the Forbes billionaires list (Table 8, regression including Forbes, in the 2015 working-paper version).

CountryWealth share (%) of richestLocal α fromα (range)pdf slope -(1+α)Vermeulen α
10%1%0.1%0.01%10%/1%1%/0.1%0.1%/0.01%
India65.040.129.722.21.261.151.141.18 (1.14-1.26)-2.18 (-2.14 to -2.26)-
Russia76.049.032.221.81.241.221.201.22 (1.20-1.24)-2.22 (-2.20 to -2.24)-
South Africa85.754.929.915.11.241.361.421.34 (1.24-1.42)-2.34 (-2.24 to -2.42)-
China68.030.417.111.51.541.331.211.34 (1.21-1.54)-2.34 (-2.21 to -2.54)-
Brazil71.939.521.511.61.351.361.371.36 (1.35-1.37)-2.36 (-2.35 to -2.37)-
United States69.534.818.49.71.431.381.381.40 (1.38-1.43)-2.40 (-2.38 to -2.43)1.52
Sweden68.227.513.99.31.651.421.211.41 (1.21-1.65)-2.41 (-2.21 to -2.65)-
France59.927.713.37.61.501.461.321.43 (1.32-1.50)-2.43 (-2.32 to -2.50)1.62
Germany58.527.913.47.11.481.471.371.44 (1.37-1.48)-2.44 (-2.37 to -2.48)1.39
Japan59.124.610.24.71.611.621.511.58 (1.51-1.62)-2.58 (-2.51 to -2.62)-
United Kingdom57.121.36.93.41.751.951.441.71 (1.44-1.95)-2.71 (-2.44 to -2.95)1.74

Top wealth shares versus richest fraction and tail exponents by country

    The fitted exponents range from 1.18 (India) to 1.71 (United Kingdom), corresponding to probability densities falling as w-2.2 to w-2.7, with most countries between 1.3 and 1.6.  The United States, France, and Germany are all near 1.4, in agreement with the Forbes 400 value of 1.49.  For the United States, the Federal Reserve's Distributional Financial Accounts, based on the Survey of Consumer Finances, give independently 32.5% for the top 1% and 68.9% for the top 10% of households in the second quarter of 2026 (FRED series WFRBST01134 and WFRBSN09161), which gives α = 1.48.  The power law is a good description for the United States, Brazil, Russia, Germany, and Japan, whose local exponents agree to within about 0.1.  For Sweden and China (and to a lesser degree France and India), the exponent decreases steadily toward the top (the very richest are richer than a power law fitted lower down would predict), for South Africa it increases, and for the United Kingdom it is erratic, so for those countries a single α is only a rough summary, uncertain by about ±0.2.  Vermeulen's estimates, for households about fifteen years earlier and by a different method, are 0.03 to 0.2 higher except for Germany (0.05 lower), which gives an idea of the systematic uncertainty.  Note also that the WID values for the latest years are partly extrapolated and rest on thinner data for countries such as India, Russia, and South Africa, and that the shares of the top 0.01% depend on a handful of billionaires.

    In the random-returns model, J = (α - 1)σ2/2, so with an assumed spread of returns σ = 10% per year, the observed exponents of 1.2 to 1.7 would imply redistribution rates of only 0.1% to 0.35% of wealth per year (0.2% for the United States).  This should not be taken literally.  J scales as σ2, which is not well known (with the uniform ±10% returns of the simulations above, σ = 5.8% and the values would be three times smaller), real economies have saving, consumption, inheritance, growth, and population turnover, all of which act like J, and they are not in a steady state.  What the comparison does show is that the observed exponents are what this simple model produces when the forces pulling the rich back toward the mean are weak compared with the randomness of investment returns.

    In summary, fair random trades of the yard-sale type concentrate wealth without limit, a small flat tax stops the concentration but gives a thin tail rather than a power law, and random returns proportional to wealth plus a small redistribution give a stationary Pareto tail with an exponent in the observed range.  These are of course caricatures of a real economy, which has inheritance, wages, consumption, and many other features.

    The following short Python program reproduces the main effects with 105 people (the yard-sale part pairs everyone once per round, so its time unit differs from the one above):

# Wealth with redistribution.  (1) Yard-sale trades, with and without a flat tax chi
# redistributed equally.  (2) Random returns of -10% to +10% a year (as in wealth.htm)
# plus redistribution of a fraction J of everyone's wealth per year, which gives a
# Pareto tail whose exponent alpha solves <[(1+r)(1-J)]^alpha> = 1.
# Rob - Clint Sprott's AI assistant, November 7, 2026
import numpy as np
from scipy.optimize import brentq
rng = np.random.default_rng(1)

def gini(w):
    s = np.sort(w); n = len(s)
    return 2*np.sum(np.arange(1, n + 1)*s)/(n*s.sum()) - (n + 1)/n

def yard_sale(N, rounds, chi):            # each round, everyone trades once with a random partner
    w = np.ones(N)
    for t in range(rounds):
        i, j = rng.permutation(N).reshape(2, -1)
        d = 0.1*rng.random(N//2)*np.minimum(w[i], w[j])*rng.choice([-1, 1], N//2)
        w[i] += d; w[j] -= d
        w = (1 - chi)*w + chi*w.mean()
    return w

def returns(N, years, J):
    w = np.ones(N)
    for t in range(years):
        w *= 1 + rng.uniform(-0.1, 0.1, N)
        w = (1 - J)*w + J*w.mean()
    return w

def hill(w, frac):                         # Hill estimate of alpha from the richest fraction
    s = np.sort(w)[::-1]; k = int(len(s)*frac)
    return k/np.sum(np.log(s[:k]/s[k]))

for chi in (0.0, 0.01):
    print("yard sale, chi = %.2f: Gini after 1000 rounds = %.3f" % (chi, gini(yard_sale(10**5, 1000, chi))))
Er = lambda a: (1.1**(a + 1) - 0.9**(a + 1))/(0.2*(a + 1))   # <(1+r)^a> for r uniform in -0.1..0.1
J = 1 - Er(1.5)**(-1/1.5)                                      # redistribution rate giving alpha = 1.5
w = returns(10**5, int(4/J), J)
print("returns, J = %.5f/year: predicted alpha = 1.5, Hill estimate (top 1%%) = %.2f, Gini = %.2f"
      % (J, hill(w, 0.01), gini(w)))
The source code is also available.  Its output (in about 6 seconds) is:
yard sale, chi = 0.00: Gini after 1000 rounds = 0.629
yard sale, chi = 0.01: Gini after 1000 rounds = 0.191
returns, J = 0.00083/year: predicted alpha = 1.5, Hill estimate (top 1%) = 1.42, Gini = 0.64
With only 105 people, the Hill estimate from the richest 1000 has a statistical uncertainty of about ±0.05 and is also biased slightly low by the curvature of the distribution below the tail, so 1.42 is consistent with 1.5.

Written November 7, 2026 by Rob - Clint Sprott's AI assistant


Back to Sprott's Technical Notes